| name | triton-generate-artifacts |
| description | Generate Triton Runner ttir, ttgir, llir, ptx, cubin, and metadata artifacts inside examples/runner/vX.Y.Z for the active Triton version and the current machine's GPU capability. Use when publishing a new Triton version's runner artifacts, regenerating checked-in example kernels, or producing compile-only artifacts for extra NVIDIA capabilities on another machine. |
Triton Generate Artifacts
Use this skill when the task is to generate or refresh versioned runner artifacts under examples/runner/vX.Y.Z.
Workflow
- Confirm the active Triton version with
python -c "import triton; print(triton.__version__)".
- Use the bundled generator script in this skill as the source of truth.
- Default to the current machine's GPU capability only.
- Generate extra capabilities only when the user explicitly asks for them.
- Keep README commands pointing at checked-in static artifacts; use this skill only to produce or refresh those files.
Commands
From the repository root:
python skills/triton-generate-artifacts/scripts/generate_runner_artifacts.py
To target explicit capabilities:
python skills/triton-generate-artifacts/scripts/generate_runner_artifacts.py 75 80 86 90 120
To target another Triton version directory without changing the environment:
python skills/triton-generate-artifacts/scripts/generate_runner_artifacts.py --version 3.7.0 75
Validation
- Verify the target version directory contains
ttir, ttgir, llir, ptx, and cubin outputs plus metadata json where applicable.
- For the local machine capability, run the corresponding README example commands when GPU execution is available.
- For extra capabilities generated via compile-only override, validate file presence and metadata
arch and triton_version instead of runtime execution.
